Skip to content
GreenKube

Sources Overview

GreenKube’s own measurements decide every recommendation. Other tools corroborate them: every source, native or third-party, implements the same contract and emits signals in one public schema, and the consensus engine builds one recommendation per target from them (Consensus across sources).

GreenKube’s measured value decides; VPA, KRR and pushed signals adjust the confidence; a source alone never creates a recommendation.

SourceKindWhat it contributesHow it is enabled
GreenKube analyzers (greenkube.rightsizing_cpu, greenkube.rightsizing_memory, greenkube.orphaned_pv, greenkube.orphaned_lb)nativeTwo-way CPU and memory rightsizing from measured usage; orphaned PersistentVolumes and LoadBalancersOn by default; each can be disabled
VPA and Goldilocks (vpa, goldilocks)open sourceContainer requests from mature VPAs in recommendation mode, including those Goldilocks createsAutomatic when the VPA CRD exists
Robusta KRR (krr)open sourceContainer requests from KRR JSON reports you pushRun KRR yourself (CI or your own CronJob) and push its report
Push API (push:<client>)customAnything your tools produce, in the public signal schemaA signals:write API key

Any other tool sends its advice through the push API. Third-party advice never reaches a pull request without going through consensus and risk review: a plugin or connector cannot create a recommendation on its own (fail closed). GreenKube’s Impact Model computes every savings figure shown; a source’s own estimate is kept for comparison only.

Karpenter, the Cluster Autoscaler and HPAs are read for the realization factor and the rightsizing evidence, not as sources.

Removed in this release

The Kubecost, GKE recommender, Azure Advisor and Polaris connectors, the chart’s KRR and Polaris CronJobs, the Karpenter NodePool heuristic, the native advisory analyzers and the consensus policies (safest, median, weighted, priority) are gone. Their settings are accepted and ignored with a warning, and POST /api/v1/sources/polaris/reports answers 404. Push the advice of those tools to POST /api/v1/signals instead. See Upgrading.

TermMeaning
SignalOne source’s proposal for one target and one capability
TargetThe cluster-qualified object a signal is about (cluster/group.Kind/namespace/name[/container])
CapabilityThe optimization domain, for example container_requests, container_limits, cleanup
ProposalThe typed change, for example container_resources or deletion; advisory is never a pull request
Engine familySet when several sources relay one engine (Goldilocks and VPA): consensus counts one vote per engine

The public schemas are signal.v1alpha1.json and signal-batch.v1alpha1.json.

Scans run in the leader controller. For each enabled source, the runner probes it (unavailable is not a failure), collects within a timeout, validates the batch (wrong source id, undeclared capability, a resource signal without a resource value, oversized payloads), and records the run and its signals. A source that keeps failing has its circuit opened for a cool-down, and its last known advice is kept. Only a complete, successful run can resolve earlier advice of that source.

The Sources page of the UI (and GET /api/v1/sources) shows every source with its status (healthy, degraded, failing, unavailable, disabled, misconfigured, pending, unregistered), last run, signal and recommendation counts, and its measured track record. Advice that never became a recommendation (no native measurement of the target) stays visible there as signals.

sources:
pluginsEnabled: true # load third-party sources installed in the image
settings: # keyed by source id
greenkube.orphaned_lb: { enabled: false }
vpa: { enabled: auto } # auto: run when the VPA CRD exists
push: { max_age_hours: 72 }

enabled is true, false or auto. An invalid entry disables that source only and is reported. Never put credentials in these settings: they live in a ConfigMap. sources.toolFloors, sources.minChange and sources.vpa.minSampleHours decide which external proposals count as votes (Configuration).

A Python package can add a source through the greenkube.sources entry point group: a Source subclass with discover(), health() and collect() that returns signals. Sources are read-only, report partial results honestly (complete=False), and pass a contract test suite (greenkube.signals.testing.check_source_contract). See docs/sources.md and the example package. If you only need to send advice, the Push API needs no code in GreenKube.